Hypergraph membrane system based [formula omitted] fully convolutional neural network for brain tumor segmentation

Abstract

Accurate segmentation is a necessary step in the clinical management of brain tumors. However, the task remains challenging due to not only large variations in the sizes and shapes of brain tumors but also wide variations among individuals. In this paper, we develop a novel fully convolutional neural network with a feature reuse module and feature conformity module (F2 FCN) to alleviate the above challenges and further improve the accuracy of segmentation. Specifically, to extract more valuable features, we present a feature reuse module to repeatedly utilize features from different layers. We also provide a feature conformity module to eliminate possible noise and enhance the fusion of different feature map levels. However, the difficult selection of multiple parameters and the long training time of a single model make CNNs less effective. To solve these problems, a new distributed and parallel computing model, a hypergraph membrane system, is designed to implement the F2 FCN. In particular, we develop a hypergraph membrane structure with three new kinds of rules to implement several F2 FCNs with different settings simultaneously to leverage the ensemble learning of F2 FCNs and save time. Experimental results on two datasets show promotive performance in terms of the Dice similarity coefficient (DSC), positive predictive value (PPV) and sensitivity compared with the state-of-the-art methods.